AM-FM Analysis of Magnetic Resonance Imagesalvarouc/papers/ulloa2013msthesis.pdf · AM-FM Analysis...
Transcript of AM-FM Analysis of Magnetic Resonance Imagesalvarouc/papers/ulloa2013msthesis.pdf · AM-FM Analysis...
AM-FM Analysis of Magnetic ResonanceImages
by
Alvaro Emilio Ulloa Cerna
Electrical Engineer, Pontificia Universidad Catolica del Peru, 2010
THESIS
Submitted in Partial Fulfillment of the
Requirements for the Degree of
Master of Science
Electrical Engineering
The University of New Mexico
Albuquerque, New Mexico
February, 2013
c©2013, Alvaro Emilio Ulloa Cerna
iii
Dedication
To my wife and son, Jessica and Gabriel, for being the main motivation of my work.
To my parents, Emilio and Norma, for their unconditional support.
To the memory of my grandmother Florencia for setting an example of dedication
and hard work for the family.
“Let your light shine before men, that they may see your good works, and glorify
your Father which is in heaven.” – Matthew 5:16
iv
Acknowledgments
I would like to thanks my advisor Dr. Pattichis for his constant guide and advice inthis thesis.
Moreover, I thank Dr. Rodriguez for his trust and encouragement to pursue agraduate degree.
Also, I would like to thanks Dr. Liu and Dr. Calhoun for their collaboration withthe data provided and discussions on interpretation of the results.
Also, thanks to the members of the Medical Image Analysis laboratory for theircomments and interest in this project. Specially to my friend Rogers for his timespent in this topic.
Finally, I express all my gratitude to my wife, Jessica, for her continuous supportand love.
v
AM-FM Analysis of Magnetic ResonanceImages
by
Alvaro Emilio Ulloa Cerna
Electrical Engineer, Pontificia Universidad Catolica del Peru, 2010
M.S., Electrical Engineering, University of New Mexico, 2013
Abstract
This thesis proposes the application of multi-dimensional Amplitude-Modulation
Frequency-Modulation (AM-FM) methods to magnetic resonance images (MRI). The
basic goal is to provide a framework for exploring non-stationary characteristics of
structural and functional MRI (sMRI and fMRI).
First, we provide a comparison framework for the most popular AM-FM methods
using different filterbank configurations that includes Gabor, Equirriple and multi-
scale directional designs. We compare the performance and robustness to Gaussian
noise using synthetic FM image examples. We show that the multi-dimensional
quasi-local method (QLM) with an equiripple filterbank gave the best results in
terms of instantaneous frequency (IF) estimation.
We then apply the best performing AM-FM method to sMRI to compute the
3D IF features. We use a t-test on the IF magnitude for each voxel to find evi-
dence of significant differences between healthy controls and patients diagnosed with
schizophrenia (n=353) can be found in the IF.
vi
We also propose the use of the instantaneous phase (IP) as a new feature for
analyzing fMRI images. Using principal component analysis and independent com-
ponent analysis on the instantaneous phase from fMRI, we built spatial maps and
identified brain regions that are biologically coherent with the task performed by the
subject. This thesis provides the first application of AM-FM models to fMRI and
sMRI.
vii
Contents
List of Figures xi
1 Introduction 1
1.1 Functional Magnetic Resonance Imaging . . . . . . . . . . . . . . . . 2
1.1.1 fMRI preprocessing . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 AM-FM demodulation methods . . . . . . . . . . . . . . . . . . . . . 4
1.2.1 Channel Component Analysis, Dominant Component Analy-
sis, and Multiscale AM-FM Analysis . . . . . . . . . . . . . . 5
1.2.2 Quasi-local n-dimensional demodulation . . . . . . . . . . . . 9
1.3 Motivation for Current Work . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 Thesis statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.5 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.6 Thesis Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2 Detection of brain activity in fMRI time-series 12
viii
Contents
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Linear Regression Analysis . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3 Independent Component Analysis . . . . . . . . . . . . . . . . . . . 15
2.4 4-D AM-FM demodulation . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4.1 Generalized quasi-local one-dimensional demodulation . . . . . 18
2.4.2 Methods comparison . . . . . . . . . . . . . . . . . . . . . . . 20
3 Preliminary results from AM-FM analysis of functional and struc-
tural MRI 25
3.1 Brain masking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2 Preliminary data analysis . . . . . . . . . . . . . . . . . . . . . . . . 26
3.2.1 AM-FM Feature space . . . . . . . . . . . . . . . . . . . . . . 28
3.2.2 Time-frequency Analysis on Resting fMRI . . . . . . . . . . . 28
3.2.3 Instantaneous frequency analysis of sMRI . . . . . . . . . . . 29
4 Results 33
4.1 Task fMRI analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.2 Resting fMRI analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.3 Structural MRI analysis . . . . . . . . . . . . . . . . . . . . . . . . . 36
5 Conclusion and future work 39
5.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
ix
Contents
5.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
A Proof that ω(x) ∈ [0, π3] when R ∈ [−∞,−1] 41
B Proof that ω(x) ∈ [2π3, π] when R ∈ [1,∞] 43
References 45
x
List of Figures
1.1 AM-FM demodulation based on Dominant Component Analysis (DCA)
and the Quasi-Eigenfunction Approximation (QEA). The input is
the extended analytic signal denoted by I(ξ). The QEA is applied
in correcting the instantaneous amplitude. This correction can be
error-prone since it depends on the estimate of the IF. . . . . . . . 8
2.1 T -map that indicates the voxels that are active for which the co-
efficients for the (a) left visual field task, and (b) the right visual
field task. The fMRI images depicts active pixels using low image
intensity. Thus, darker pixels signify activation. . . . . . . . . . . . 15
2.2 ICA components coherent with visuo-motor cortical regions and time
courses for both tasks that the patient perform. . . . . . . . . . . . 17
2.3 θ(R) and θ(−R) in the respective domain R. . . . . . . . . . . . . . 20
2.4 Frequency Discrimination Algorithm (FDA) block diagram. Where
demod. 1 represents eq. (1.3), demod. 2 represents eq. (1.5), hLP
and hHP are the low pass and the high pass filter respectively. . . . 21
xi
List of Figures
2.5 AM-FM analysis filter-banks. (a) Directional Filter Bank (DFB),
each band indexed from 0 to 7; (b) Block diagram of (c) Nonsub-
sampled directional multiscale and multiresolution filterbank block
diagram. [1] and (d) the Pyramidal Directional Filter Bank (PDFB)
with a multiresolution and multiscale decomposition into octave bands
and a directional decomposition by the DFB to each high-pass band
increasing the number of bands with the scale. Adapted from [2]. . 22
2.6 (a) Ichirp with frequencies ω1,2 ∈ [0, π] (see eq. (2.3)) and (b) Icubic
with x0 = −6 and x1 = 6 (see eq. (2.4)) . . . . . . . . . . . . . . . . 22
2.7 A single line of pixels showing the instantaneous frequency magnitude
estimate for (a) QLM, and (b) QEA on Ichirp with gaussian noise. . 24
2.8 Normalized error comparisons for different filter-banks. (a) Filter-
bank comparison using QEA demodulation method, and (b) filter-
bank comparison using generalized QLM demodulation method. Er-
ror bars denote the minimum and maximum error for 50 repetitions
at the given noise level. . . . . . . . . . . . . . . . . . . . . . . . . 24
3.1 (a) Histogram of the fMRI data, (b) slice 14 of the mask and (c)
masked fMRI. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.2 Matrix plot of AM-FM features and BOLD fMRI for raw data (upper
plot) and after edge and eye blinking artifacts have been removed
(lower plot). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
xii
List of Figures
3.3 (a) Manually selected Principal component showing coherence with
brain function using the feature space containing AM-FM functions
and (b) coefficients of the component corresponding to IP and ex-
pected response (HRF) of the BOLD signal for the task exiciing the
right visual and motor cortices. . . . . . . . . . . . . . . . . . . . . 32
3.4 (a) Scatter plot of IFS and IFt and (b) bivariate histogram for one
healthy subject. A logarithm transformation was applied for the
probability frequencies in each bin to enhance visualization. . . . . . 32
4.1 Instantaneous phase of axial slice 11 (left) and slice 24 (right) to show
activation at visual and motor areas in the brain for a (a) subject in
rest, (b) performing left task and (c) performing right task. . . . . . 34
4.2 (a) Histogram of instantaneous phase and (b) Histogram of BOLD
signal. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.3 Biologically relevant ICA on instantaneous phase components for (a)
left task and (b) right task. . . . . . . . . . . . . . . . . . . . . . . . 36
4.4 (a) P-value map for each bin tested and (b) Bins that passed bonfer-
roni correction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.5 Log10(P ) thresholded at 5 with a T1 image as template. . . . . . . 38
xiii
Chapter 1
Introduction
Non-stationary image processing methods provide a powerful alternative to station-
ary approaches. Non-stationary analysis can provide significant new insight into the
basic signal characteristics. In this Master’s thesis, we explore non-stationarity ex-
hibited in functional magnetic resonance imaging (fMRI) and structural magnetic
resonance imaging (sMRI). It is known that MRI signals [3] and noise [4, 5] exhibit
strong non-stationary characteristics.
For one-dimensional signals, the short-time Fourier transform (STFT) is a stan-
dard tool for exploring non-stationary characteristics. More generally, one-dimensional
time-frequency analysis requires the use of the Wigner distribution and its derivatives
[6]. One-dimensional Wavelet transforms provide a specific time-frequency analysis
method [7, 6].
The Teager-Kaiser operator in [8, 9] provided an alternative method for studying
one-dimensional non-stationary signals based on the use of amplitude modulated
and frequency modulated (AM-FM) functions. This was further extended to two-
dimensional AM-FM representations for digital images in [10] and [11]. Moreover,
multi-dimensional FM modeling for non-stationary image structure was studied in
1
Chapter 1. Introduction
[12]. Recently, Dr. Murray introduced the use of multi-scale AM-FM decompositions
in [13].
This introductory chapter is organized as follows. In sec. 1.1, we provide an
general description of acquisition and pre-processing of fMRI data. In sec. 1.2,
we provide an introduction to some of the basics of AM-FM models and AM-FM
demodulation.
1.1 Functional Magnetic Resonance Imaging
fMRI is a non-invasive technique that provides an image of brain function through
time [14]. This imaging technique gives information on blood oxygenation indirectly
related to neuronal activity. The blood oxygen level dependent (BOLD) contrast
changes trough time as a series of three dimensional images collected while the subject
may be receiving stimulus, performing a task or resting.
The MRI scanner induces a magnetic field that causes the protons to align their
spins. Then, by introducing a pulse of magnetic energy perpendicular to the main
magnetic field in the form of a radio frequency pulse (specific to the hydrogen atom
due to its high concentration in the human body), the spins absorb energy and be-
come excited [15]. The time it takes for the protons to return to their equilibrium
magnetization is an exponential decay process with time constant parameters called
transverse relaxation time, T2, and longitudinal relaxation time, T1. Also, the param-
eter T ∗2 is a time constant that characterizes the exponential decay of the signal, due
to spin-spin interactions (related to T2), magnetic field inhomogeneities (magnetic
field induced by the MRI machine), and susceptibility effects (physical property of
the brain). The rapid decay of the T ∗2 signal provides for a faster MRI scanning.
The time sampling interval, TR, is designed to be long enough so that gray matter
2
Chapter 1. Introduction
tissue can fully recover in between pulses. Depending on the tissue of interest,
parameters can be tuned to enhance its contrast with surrounding tissue.
The MRI signal is acquired from two orthogonal detectors which are put in com-
plex form [16], where BOLD contrast is measured. The BOLD signal depends on
the magnetic properties of blood carrying oxygen to the neurons (oxygenated) and
residual blood after neurons have used the oxygen (de-oxygenated). Thus, BOLD
indirectly indicates neuronal activity.
1.1.1 fMRI preprocessing
Following scanning, the fMRI data is pre-processed to enhance data quality and allow
for improved statistical power. The pre-processing steps are performed using SPM 8
software as described in [17]. The basic steps include:
Step 1. Slice Timing
The slices are collected at different times and require synchronization to avoid
signal biases. Slice timing correction is performed using sinc function interpo-
lation.
Step 2. Realignment
Head motion during scanning, even in the order of millimeters, can still cause
significant artifacts. Motion correction is achieved realigning a time-series of
images acquired from the same subject using a least squares approach and a
6 parameter (rigid body) spatial transformation. Also, the fat chemical that
envelops the brain causes a shift in the resulting image and the susceptibility
map is not homogenous due to the air canals near the brain such as the audi-
tory and nasal canals. This correction is performed using an estimate of the
susceptibility map and reconstruction from the phase of the acquired image.
3
Chapter 1. Introduction
Step 3. Spatial Normalization
Spatial normalization involves image registration to the brain atlas to allow for
comparisons among different individuals.
Step 4. Smoothing
Spatial smoothing is applied for improving the signal-to-noise ratio (SNR) to
allow for better activation detection. Smoothing does not have a high impact on
frequency estimation because it is only reducing amplitude and not distorting
frequency content.
1.2 AM-FM demodulation methods
AM-FM demodulation methods are in constant growth and several applications have
been presented. The list of application includes fingerprint classification [18], image
segmentation [19, 20, 21], and ultrasound video analysis [22, 23, 24, 25], among
others.
The AM-FM representation for multi-dimensional signals provides an effective
model of non-stationary content [12]. For the AM-FM representation, a multi-
dimensional signal I(x) is expressed as a sum of AM-FM components using:
I(x) 'M∑m=1
am(x) cos(ϕm(x)) (1.1)
where x represents a vector of spatio-temporal coordinates, I(x) : Rk → R maps the
voxel values for each coordinate x, M is the number of components, an > 0 denote the
instantaneous amplitude (IA) components, and ϕn denote the instantaneous phase
(IP) components. Slowly-varying (amplitude) characteristics are modeled using the
IA components. Significant changes in non-stationary frequency content are modeled
using the IP components.
4
Chapter 1. Introduction
Given the multidimensional signal I(x), the computation of the AM-FM com-
ponents involves the estimation of the instantaneous amplitude functions an(x), the
instantaneous phase functions ϕn(x), and the instantaneous frequency (IF) function
defined as the gradient of the instantaneous phase:
ωn(x) = ∇ϕ(x)
=
(∂
∂x1ϕn(x),
∂
∂x2ϕn(x), . . . ,
∂
∂xkϕn(x)
).
1.2.1 Channel Component Analysis, Dominant Component
Analysis, and Multiscale AM-FM Analysis
In this section, we introduce the most popular methods for deriving AM-FM models.
In order of complexity, the simplest method is based on Channel Component Analysis
(CCA), followed by Dominant Component Analysis (DCA) [10], and multi-scale AM-
FM analysis [13].
CCA consists of the decomposition of the input signal into M components com-
ing from M channels. DCA constructs a single AM-FM component by using the
channel output with the highest IA estimate. The Multi-scale analysis applies DCA
to different frequency bands.
AM-FM models use AM-FM demodulation methods in order to estimate the
AM-FM functions IA and IP. With the exception of the quasi-local method (QLM),
all other methods require the computation of the extended analytic signal. For a
single AM-FM component, the extended analytic signal attempts to estimate a com-
plex exponential a(x) exp(jϕ(x)) from the real-valued input AM-FM component of
a(x) cos(ϕ(x)). In multiple-dimensions, the extended analytic signal representation
is approximated using three steps:
Step 1. Take the multidimensional Fast Fourier Transform of the input signal I(.).
5
Chapter 1. Introduction
F(ω) = FFTN(I(x))
Step 2. Remove half of the spectrum by setting the components to zero and dou-
bling the remaining components:
G(ω) =
0, if ω1 < 0
2F(ω), otherwise.
Step 3. Take the inverse FFT to estimate the complex signal approximation.
IAS(x) = IFFTN(G(ω))
Note that no signal information is lost trough the computation of the extended
analytic signal due to conjugate symmetry of the multidimensional Fourier Transform
(F(ω) = F∗(−ω)) in real-valued signals.
The extended analytic signal IAS is then processed through a filter-bank. This
procedure is outlined in Fig. 1.1, where the impulse responses of the filter-bank are
given by: h1, h2, . . . , hN . In the basic channel component analysis approach, the
outputs from each filter are used to estimate the AM-FM components. This approach
is clearly best if non-stationary components can be localized to specific bands that
is represented in the left figure of Fig. 1.1.
The assumption that the AM-FM components cannot be constrained to specific
frequency bands leads to DCA. In DCA, each AM-FM component an(x), ϕn(x) and
ωn(x) are computed at each voxel as shown in the right figure of Fig. 1.1. Then, the
goal of the filterbank is to provide a voxel-based separation of the different AM-FM
components. AM-FM components are extracted from different filters at each voxel.
In DCA, the dominant AM-FM component is selected by the channel that gives
the largest IA estimate. Due to the need to capture non-stationary behavior, the
6
Chapter 1. Introduction
dominant filter can vary from voxel to voxel. Formally, we have the dominant IA
determined using:
a(x) = maxn∈[1, M ]
{an(x)},
and the index of the maximum channel
κ(x) = argmaxn∈[1, M ]
{an(x)}
is then used in determining the instantaneous phase and instantaneous frequency
estimates using
ϕ(x) = ϕκ(x)(x), and
ω(x) = ∇ϕκ(x)(x).
The instantaneous frequency based on the QEA is estimated by:
ω1 = cos−1
[IAS(x1 + 1, x2, x3, x4) + IAS(x1 − 1, x2, x3, x4)
2IAS(x)
]
ω2 = cos−1
[IAS(x1, x2 + 1, x3, x4) + IAS(x1, x2 − 1, x3, x4)
2IAS(x)
]
ω3 = cos−1
[IAS(x1, x2, x3 + 1, x4) + IAS(x1, x2, x3 − 1, x4)
2IAS(x)
]
ω4 = cos−1
[IAS(x1, x2, x3, x4 + 1) + IAS(x1, x2, x3, x4 − 1)
2IAS(x)
]
An extension of DCA, called multi-scale analysis, was proposed by Dr. Murray in
his dissertation [13]. In multi-scale analysis, we simply apply dominant component
analysis over different collections of bandpass filters. The basic idea is to group
together the different channel filters based on frequency magnitude. Thus, we can
have low, medium, and high frequency scales. Furthermore, multi-scale analysis can
take advantage of combinations of scales.
7
Chapter 1. Introduction
Figure 1.1: AM-FM demodulation based on Dominant Component Analysis (DCA)and the Quasi-Eigenfunction Approximation (QEA). The input is the extended an-alytic signal denoted by I(ξ). The QEA is applied in correcting the instantaneousamplitude. This correction can be error-prone since it depends on the estimate ofthe IF.
The selection and design of an appropriate filter-bank can significantly affect
the resulting AM-FM decomposition. Ideally, filterbank design should reflect the
properties of the signals that are being analyzed. General-purpose Gabor-based
filter-banks were used in Pattichis et al. [21] and Havlicek et al. [26]. Separable,
flat-passband filters were used by Rodriguez in [24]. More recently, Dr. Murray [27]
implemented a dyadic, separable and multiscale filter-bank using equiripple filters.
Given the fact that the filter-bank can significantly impact the estimated AM-FM
components, in Chapter 3, I compare accuracy between estimates derived using both
Gabor and equiripple filter-bank designs.
8
Chapter 1. Introduction
1.2.2 Quasi-local n-dimensional demodulation
For the quasi-local method, estimation is based on products of samples of AM-FM
components. We begin by defining
g(ε1,ε2) = I(x+ ε1)I(x− ε2),
where (ε1, ε2 ≥ 0), and I(x) = a(x) cosϕ(x). Then, the instantaneous amplitude
can be estimated using
a(x) =√
2g(0,0)(x) (1.2)
where g(ε1,ε2)(x) = hLP (x) ∗ g(ε1,ε2)(x) and hLP (x) is a low pass filter. The low-pass
filter is used for removing second-order, higher-frequency terms. For the purposes
of this research, the low-pass filter was designed using an equiripple design with
frequency cut-off set to π/10, a transition band of π/10, passband ripple at 0.017
dB, stop-band attenuation of 66.02 dB as described in [13, sec. 3.3].
Assuming that the instantaneous frequency satisfies ω(x) < π/2, it can then be
estimated using
ω(x) = cos−1
(R(x) +
√R2(x) + 8
4
)(1.3)
where
R(x) =2g(1,1)(x)
g(1,0)(x) + g(0,1)(x). (1.4)
This operator is separable and applied to every dimension of the data. For π/2 <
ω(x) < π, the instantaneous frequency is estimated as proposed in [24]
ω(x) = π − cos−1
(−R(x) +
√R2(x) + 8
4
)(1.5)
9
Chapter 1. Introduction
1.3 Motivation for Current Work
Classical MRI analysis methods either assume stationarity or ignore non-stationary
characteristics. The goal of this thesis is to apply AM-FM methods that are well-
suited for analysing non-stationary signals to functional and structural MRI.
AM-FM methods provide a rich representation which allows the analysis in terms
of the instantaneous amplitude, frequency, and phase for each voxel of the image.
AM-FM functions can adapt to significant signal changes.
The thesis will explore the application of AM-FM methods to model brain dy-
namics and signal coherence in MRI images, providing a new data representation
to compare between patients with mental diseases and healthy controls. The focus
of the thesis is on the statistical analysis and interpretation of AM-FM components
extracted from functional and structural MRI.
1.4 Thesis statement
The primary thesis of this dissertation is that the analysis of MRI data by AM-FM
methods can lead to new insights into the non-stationary characteristics of structural
and functional MRI.
First, we provide a comparison framework for the most popular AM-FM ap-
proaches to select an optimal method based on its performance in resistance to noise.
Second, we apply the best-performing AM-FM demodulation method to functional
and structural MRI where we explore its results with statistical tools. We compare
sMRI instantaneous frequency in a case-control study using two independent sample
t-test and logistic regression. Also, we use principal component analysis and inde-
pendent component analysis on the instantaneous phase to describe its properties on
10
Chapter 1. Introduction
fMRI.
1.5 Contributions
The main contributions of the present work include:
• An AM-FM analysis framework for comparing demodulation methods (QEA
and QLM) based on 3 filterbanks (Gabor, Equiriple and Multiscale directional).
• A statistical analysis of MRI signals using a time-frequency non-stationary
representation.
• An extended application of the instantaneous phase for detecting brain activity
in functional images, and the use of the instantaneous frequency for represent-
ing brain structure.
• Provide evidence of significant differences between patients and healthy-controls
given by AM-FM analysis that cannot be seen analyzing the raw data.
1.6 Thesis Overview
The thesis is organized into four parts. The first chapter presented a brief review
of MRI data acquisition and AM-FM methods. The second chapter summarizes the
classical methods used in brain activity detection and introduces the comparison
framework. The third chapter describes the application of AM-FM methods to MRI
and provides suggestions for differentiating between patients and healthy controls.
Chapter four provides a discussion of the results. Finally, in the last chapter we
provide our final conclusions based on the results and suggest future work.
11
Chapter 2
Detection of brain activity in
fMRI time-series
2.1 Introduction
Despite the development of several approaches, the detection of true brain activity
in fMRI scans remains an open problem. In this chapter, I summarize two of the
most popular approaches.
The first approach is based on massive univariate statistical inference using a
linear regression model. The BOLD signal is set as the response variable and task
designed responses are set as the explanatory variables. These designed responses
for each task are generated with the HRF convolved with a box-car (one for stimuli
presence). Then, the model is fit for each voxel independently [28].
It is important to consider the multiple correction problem since we used a sta-
tistical test for determining whether each voxel is linearly related of the designed
BOLD response or not. The probability of rejecting the null hypotheses H0 : β = 0
12
Chapter 2. Detection of brain activity in fMRI time-series
when it is true (false rejection), increases with the number of tests. Thus, we may
have a large probability of making one or more false rejections of the true H0. The
most conservative way to correct the p-values is using Bonferroni p-value correction
for voxels inside the brain only.
Multivariate approaches can help model brain connectivity represented by tem-
poral correlations among spatially separated brain regions. A popular multivariate
method for fMRI is based on Independent Component Analysis (ICA). ICA has
been applied for the identification of various signal-types in the spatial or temporal
domain, second level analysis of fMRI data and for the analysis of complex-valued
fMRI data [29]. ICA can be applied without the need to have a-priori models for
brain activity and produce spatially independent components (component images)
or temporal independent components (component time courses).
In the remaining part of the Chapter, we introduce the use of the two most pop-
ular methods for brain analysis for fMRI data: linear regression analysis in Section
2.2 and independent component analysis (ICA) in Section 2.3. Also, we propose an
adaptation of the AM-FM analysis methods for analysing MRI images is given in
Section 2.4.
2.2 Linear Regression Analysis
Statistical Parametric Mapping (SPM) is a popular software tool to construct and
assess spatial statistical processes and test hypotheses about functional imaging data.
In SPM, we specify each voxel time series as the response variable and, in case of a
block design, the convolution of the stimulus function of experiment with the HRF
as predictor variable.
Then, the estimated coefficients for each voxel or so called β maps show regions of
13
Chapter 2. Detection of brain activity in fMRI time-series
the brain in which the model respond to the predictors. SPM also provides T -values
resulting of hypothesis testing on β with the estimated standard deviation associated
with this coefficient. This allows a visual inspections of regions of the brain linearly
related to the stimuli and analysis with previous anatomic and functional knowledge
of brain.
Example T -maps are shown in Fig. (2.1). In the example, we have the fMRI scan
of 1 subject performing a visual and motor task [30]. The task consists of a flickering
checker board presented to the subject as a visual stimuli. This is presented for 15
seconds to the right side and it is followed by 5 seconds of rest to the subject. Next,
the same visual stimuli are presented to the left side of the visual field for 15 seconds
and finally 20 seconds rest to repeat of the same pattern four times. For simplicity,
we show the T -map using as explanatory variables the two designed BOLD response
that relates to each side stimuli. The T -map indicates with black values the voxels for
which the coefficients significantly different from zero. Notice that motor and visual
cortex are those with the highest T statistic. Since this is a univariate method and
each voxel is tested independently, we note sparse isolated regions on the T -map that
are not related to the stimuli. The main issue with this method is that we require
a prior model of brain behaviour and a model of the BOLD response which may
vary by subject. Furthermore, this is a massive univariate approach which requires
correction of the observed significance by the number of comparisons which is equal
to the number of intro-cranial voxels (∼ 30, 000).
On the other hand, as we shall discuss next, multivariate approaches such as
PCA and ICA work on whole images and can identify spatio-temporal patterns over
voxels without the need to specify priori information.
14
Chapter 2. Detection of brain activity in fMRI time-series
(a) (b)
Figure 2.1: T -map that indicates the voxels that are active for which the coefficientsfor the (a) left visual field task, and (b) the right visual field task. The fMRI imagesdepicts active pixels using low image intensity. Thus, darker pixels signify activation.
2.3 Independent Component Analysis
Independent component analysis is a multivariate method that separate a multivari-
ate signal into additive components that are statistically independent. ICA seems
to capture the essential structure of the data in many applications, including feature
extraction and blind signal separation [31]. ICA find this components assuming that
each observed signal is a linear mixture of n independent components.
In matrix form, we have X = AS where X are the observed signals, A is the
mixing matrix and S are the sources. Thus, the problem is to estimate A and S
from X such that each source in S is as independent as possible. From X = AS, we
have that each observation is represented as
xj = aj1s1 + aj2s2 + · · ·+ ajnsn. (2.1)
ICA cannot determine the variances of the independent components since any
15
Chapter 2. Detection of brain activity in fMRI time-series
scalar multiplier in A can be cancelled by a division in S. Also, the method can not
determine the order of the independent components.
In addition, Gaussian sources are not separable in the ICA sense [32] because
any orthogonal transformation will have the exact same joint distribution and thus
we cannot estimate the mixing matrix A. To avoid this problem, we measure the
kurtosis [33] and the neg-entropy [34] to determine whether an R.V. exhibits a normal
distribution.
In order to estimate the sources, we first decorrelate the observation signals using
a whitening matrix W . Higher order dependence can be partially addressed through
rotations so as to minimize the dependence of the estimated sources. To minimize
dependence, estimation proceeds by estimating A, W that minimize mutual infor-
mation of the resulting source signals [35]. Similarly, A, W can be selected based on
minimizing neg-entropy or maximizing the output entropy [36] (INFOMAX) among
the sources. Alternatively, estimation can be performed using maximum likelihood
estimation [37] among others.
ICA has become a very popular tool for fMRI analysis [29], mainly due to the
consistency of its results with a priori biological knowledge regarding brain function.
The assumption of ICA on fMRI is that activated regions of the brain do not overlap
in time or space. Therefore, regions of the brain are independent sources that mix
linearly and generate the observed fMRI data [38]. Also, it has been shown that
fMRI noise is not Gaussian. In practice, most methods use the amplitude of the
complex fMRI signal where the Gaussian noise of the real and imaginary part is
transformed to Rician noise.
As an example, we apply ICA using the same data used in sec. 2.2 and show
results using the software tool Group ICA Of fMRI Toolbox (GIFT) [39]. In this case,
15 independent sources were selected. In Fig. 2.2, we only show the two components
16
Chapter 2. Detection of brain activity in fMRI time-series
10 20 30 40 50 60 70 80 90 100 110
IC 8
IC 9
Figure 2.2: ICA components coherent with visuo-motor cortical regions and timecourses for both tasks that the patient perform.
that are relevant to the task. We notice activation regions that extend beyond the
visual and motor cortices. The results are consistent with the task and the Linear
Regression results of Fig. 2.1.
2.4 4-D AM-FM demodulation
To the best of our knowledge, this is the first application of AM-FM methods to
fMRI data. Furthermore, this is the first application of AM-FM models to 4D data
sets. The application requires the design of a four dimensional filter-bank. In this
application, we will consider the use of two AM-FM scales to avoid expensive com-
putation.
17
Chapter 2. Detection of brain activity in fMRI time-series
The rest of this section is organized as follows. First, we provide details on
the development of a new quasi-local method in section 2.4.1. Then, we evaluate
performance on a synthetic example using three different filter-banks.
2.4.1 Generalized quasi-local one-dimensional demodulation
In this section, we propose a modification of the Quasi-local method to work with
arbitrary frequency bands. The standard quasi-local method constrains the signal
to either [0, π2] or [π
2, π] and this constrains the design of the associated filter-bank.
Therefore, we propose this new method in order to allow for generic filter-bank
designs. Since the quasi-local approach is separable, we develop the one-dimensional
algorithm that is applied along each dimension.
Let the monocomponent modulated signal be f(x) = a(x) cosϕ(x), where a(x) is
the instantaneous amplitude and ϕ(x) is the instantaneous phase. Also, let ω(x) be
the instantaneous frequency of f(x) computed with
ω(x) = cos−1
(±R(x) +
√R2(x) + 8
4
)where
R(x) =2g(1,1)(x)
g(1,0)(x) + g(0,1)(x),
g(ε1,ε2) = f(x+ ε1)f(x− ε2),
and ε1, ε2 ≥ 0.
Then we know from sec. 1.2.2 that the problem is divided into two cases: ω(x) <
π/2 and π/2 < ω(x) < π for which we either apply eq. (1.3) or eq. (1.5) respectively.
We use a low-pass and high-pass filter to get two versions of f(x) which are
f1(x) = hLP (x) ∗ f(x)
f2(x) = hHP (x) ∗ f(x)
18
Chapter 2. Detection of brain activity in fMRI time-series
where hLP has a passband of 0 < ω < π/2, and hHP has a passband of π/2 < ω < π.
Now, we define
θ(R) =R +√R2 + 8
4(2.2)
where θ(R) is the argument of cos−1(.) in eq. (1.3) and θ(−R) becomes the argument
of cos−1(.) in eq. (1.5).
Now, note that
0 < cos−1(θ(R)) <π
2
, and
π
2< π − cos−1(θ(−R)) < π
Then R ∈ (−∞, 1] and −R ∈ [−1,+∞).
Now, I plot the values of θ(R) and θ(−R) in Fig. 2.3, where we observe that
θ(R) is monotonically increasing so 0 < θ(R) < 0.5 when R ∈ (−∞,−1]. Using eq.
(1.3) we can establish that 0 < ω(x) < π3. Therefore, we use eq. (1.3) to estimate
ω(x) (see proof in Appendix A).
Also, notice that θ(−R) is monotonically decreasing so 0 < θ(R) < 0.5 when
R ∈ [1,∞). Using eq. (1.5) we can establish that 2π3< ω(x) < π. Therefore, we use
eq. (1.5) to estimate ω(x) ( see proof in Appendix B). Thus, we can choose what
demodulation equation to use according to the computed value of R which is for
ω(x) ∈ [0, π3] ∪ [2/pi
3, π]. Finally, DCA is applied to select between f1(x) and f2(x)
for ω(x) ∈ [π3, 2π
3] is computed and merged with the previous outcome. This process
is outlined as a diagram block in Fig. 2.4.
19
Chapter 2. Detection of brain activity in fMRI time-series
Figure 2.3: θ(R) and θ(−R) in the respective domain R.
2.4.2 Methods comparison
Filter banks
In order to have the most representative filter banks used for AM-FM demodulation
we test three filter bank designs: (i) non separable Gabor filter bank [40, 10], (ii)
separable filters designed by least squares optimization used in [24], (iii) and the
directional multiscale and multiresolution filter bank of [1]. The Gabor filter bank
is described in [10, sec. 4.4] and has been applied in [41, 42, 43]. The Gabor
filter bank has also been used in models of the human visual system (e.g. see [10,
sec. 2.2]). Separable filter banks designed based on least squares approximation
in [24, sec. 3.2.8.1] provides for a direct approach. The directional multiscale and
multiresolution filter-bank [2] provides an alternative approach.
In [2], the author states that a directional filter bank is made to support high
frequency components (representing directionality) of images, so, low frequency com-
ponents are handled poorly. Thus, in an effort to improve this problem, a multiresolu-
tion analysis is considered where low frequencies are partitioned with less directional
20
Chapter 2. Detection of brain activity in fMRI time-series
Figure 2.4: Frequency Discrimination Algorithm (FDA) block diagram. Where de-mod. 1 represents eq. (1.3), demod. 2 represents eq. (1.5), hLP and hHP are thelow pass and the high pass filter respectively.
bands (see Figure 2.5d). The second approach is also used in the Nonsubsampled
Contourlet Transform (NSCT) implemented in [1] which preserves the image size in
each scale. According to [1], the NSCT results with better frequency selectivity and
regularity when compared to the SCT.
Synthetic Frequency Modulation Images
We test the use of different filter-banks using two synthetic FM images. In particular,
we test demodulation using quadratic and cubic phase examples.
The first synthetic example is based on a chirp image using:
Ichirp(x1, x2) = cos
[1
2N(ax21 + bx22)
](2.3)
21
Chapter 2. Detection of brain activity in fMRI time-series
(a) (b) (c) (d)
Figure 2.5: AM-FM analysis filter-banks. (a) Directional Filter Bank (DFB), eachband indexed from 0 to 7; (b) Block diagram of (c) Nonsubsampled directionalmultiscale and multiresolution filterbank block diagram. [1] and (d) the PyramidalDirectional Filter Bank (PDFB) with a multiresolution and multiscale decompositioninto octave bands and a directional decomposition by the DFB to each high-pass bandincreasing the number of bands with the scale. Adapted from [2].
where N is the number of pixels by side of the square image. The instantaneous
frequency is set to vary from −π to π in each direction. The resulting synthetic
image is shown in Fig. 2.6a.
(a) (b)
Figure 2.6: (a) Ichirp with frequencies ω1,2 ∈ [0, π] (see eq. (2.3)) and (b) Icubic withx0 = −6 and x1 = 6 (see eq. (2.4))
22
Chapter 2. Detection of brain activity in fMRI time-series
The second synthetic example is generation using:
Icubic(x1, x2) = cos(x31 + x32 − ax1 − bx2
)(2.4)
where, the instantaneous phase is ϕ(x) = (x31+x32+3ax1−3bx2), so the instantaneous
frequency is:
∇ϕ(x) = ω(x) = (∂ϕ
∂x1,∂ϕ
∂x2) = (3x21 + 3a, 3x22 − 3b).
We set a = −6 and b = 6 to achieve a maximum frequency to 0.3476 cycles/pixel as
shown in the image in Fig. 2.6b. All images are of size 512× 512 pixels.
Results
We measure the error in the instantaneous frequency using
error =||ωt − ωest||2||ωt||2
(2.5)
where ωt is the true value and ωest is the estimated value. We use the term normalized
error to refer to (2.5).
QLM estimation is not intended for use for very low or very high frequencies. We
thus examine the error in the normalized frequency interval of 0.01π ≤ |ω| ≤ 0.99π.
We also do not consider strong boundary artifacts by only measuring errors that are
at-least 5 pixels away from the boundary of the image. In other words we do not
consider the error in the upper, lower, left, and right edges of the image (see Fig.
2.7).
For final results in Fig. 2.8, we take the mean between the errors on Ichirp and
Icubic and report the error for each level of noise. The lowest error is given by the
equiriple filterbank using the QLM demodulation method. This is the method that
will be used for further analysis.
23
Chapter 2. Detection of brain activity in fMRI time-series
Figure 2.7: A single line of pixels showing the instantaneous frequency magnitudeestimate for (a) QLM, and (b) QEA on Ichirp with gaussian noise.
(a) (b)
Figure 2.8: Normalized error comparisons for different filter-banks. (a) Filter-bankcomparison using QEA demodulation method, and (b) filterbank comparison us-ing generalized QLM demodulation method. Error bars denote the minimum andmaximum error for 50 repetitions at the given noise level.
24
Chapter 3
Preliminary results from AM-FM
analysis of functional and
structural MRI
In this chapter, we explore AM-FM demodulation applied to sMRI and fMRI. The
goal is to explore non-stationary behaviour in MRI using the extracted AM-FM
features.
The outline of this chapter is as follows. In sec. 3.1, we describe the method
for masking out data outside of the brain. In sec. 3.2, we explain our exploratory
analysis and discuss preliminary results.
3.1 Brain masking
There are voxels that fall outside of the brain and are not of interest for the study.
The boundary between intra-cranial voxels and background is fuzzy. In part, this is
25
Chapter 3. Preliminary results from AM-FM analysis of functional and structural MRI
(a) (b) (c)
Figure 3.1: (a) Histogram of the fMRI data, (b) slice 14 of the mask and (c) maskedfMRI.
due to the smoothing in the pre-processing steps. Therefore, we require a threshold
to separate intra-cranial voxels from background.
A common method is to set the threshold to the mean value. The mean value
is usually a valid threshold since around half of the total number of voxels are out
of the brain. A more general approach for two dimensional images is proposed in
[44]. It suggests to take the histogram of the image, smooth it and find the local
minimum that discriminates the object from the background, assuming a bimodal
distribution (see Fig 3.1). We extend this idea to the four dimensional data (fMRI).
This approach provides a more general image segmentation for fMRI data not being
dependent on the number of voxels covered by the brain in the four-dimensional
data. In what follows, the analysis is applied only to voxels that are segmented as
inside the cranium. See Fig. 3.1.
3.2 Preliminary data analysis
First, we focus on fMRI data collected from an experiment designed to stimulate
visual and motor cortices, acquired with an echo planar sequence for a period of 3
26
Chapter 3. Preliminary results from AM-FM analysis of functional and structural MRI
minutes and 40 seconds, which gives a total of 220 time points with TR set to 1
second. See [30] for complete detail about acquisition and pre-processing.
The experiment is composed of three different task sequences. First, the subject
is exposed only to a visual stimuli. Second, the subject performs to a motor task
(finger tapping). Third, the subject is exposed to a visual stimuli while performs a
motor task. We focus our study to the third sequence so the data set is reduced to
124 time points.
We compute IA, IF and IP estimations using QLM with an equiripple filter-
bank designed with 17 coefficients, 2 scales, maximum amplitude ripple of 0.02 in
the pass band, 0.2 maximum amplitude ripple in the rejection band and a 10%
frequency spacing for the transitions. For filtering, we design a separable filter of
17× 17× 17× 17 coefficients based on one 1D equiripple design as outlined earlier.
The original dataset is of size 53× 63× 46× 124. We discard the first 7 time points
as in [30] and constrain the analysis to 117 time points where the subject performs
a visual and motor task simultaneously.
For the exploration of the IF, IP and IA of the fMRI, we show a matrix plot that
allows visual inspection of relationships between these AM-FM features and BOLD
signals (see Fig. 3.2.a). From the plot, we notice that some values of the instanta-
neous frequency on time (IFt) are well-separated from the primary distribution and
appear to be outliers. A reconstruction of the location of these values in voxel space
indicates that the outlying values are primarily due to eye blinking and, thus, not
useful for our analysis. Blinking causes rapid changes in time and are not of interest.
In addition, the IP values close to −π and +π were found to be associated with
the boundary between the brain and the background. Thus, these points were also
removed. See Fig. 3.2.b for the data without blink eye effect and boundary effects.
27
Chapter 3. Preliminary results from AM-FM analysis of functional and structural MRI
3.2.1 AM-FM Feature space
In order to explore AM-FM features and their relation to brain function, first, we
normalize the features to have zero mean and unit standard deviation across time.
Then, we define a feature matrix composed by the concatenation of all AM-FM
features along the time dimension, resulting in a matrix of 117×5 time points by the
number of voxels. Each feature has 117 time points that are concatenated to form
the feature space in the following order
[ω1, ω2, ω3, ωt, ϕξ].
Next, we reduce the number of time points using a Principal Component Analysis
(PCA) to 92 features that keeps 95% of the total variance. Finally, we take the PCA
components and display them as spatial maps. Visual inspection of each spatial map
reveals that one of these components highlights the areas of the visual and motor
cortices. Then, we examine the coefficients corresponding to this component and
find evidence that the IP feature is related to the task (See Fig. 3.3). Notice that
IP presents temporal coherence with the visuo-motor task given to the subject. This
motivates further exploration of the IP and it relation to brain activity since, based
on our observations,this is the most informative feature for task based fMRI.
3.2.2 Time-frequency Analysis on Resting fMRI
Resting-state fMRI is a functional scan of a resting subject and is usually performed
to detect abnormal activity in patients using a functional conectivity model [45]. For
instance, many studies on depression [46], autism [47], attention deficit hyperactivity
disorder [48] and schizophrenia [49] (among others) have used resting-state fMRI.
Several approaches have been proposed for the analysis of resting fMRI [50, 51, 52]
and a baseline study is proposed in [53], which, among other results, revealed strong
28
Chapter 3. Preliminary results from AM-FM analysis of functional and structural MRI
effects of age and gender in the power spectrum of certain regions in the brain.
Moreover, frequency content has proven to be useful in the analysis of resting-state
fMRI [54, 55]. However, a time-frequency analysis such as AM-FM demodulation
has not been applied yet.
We therefore propose a scheme to process resting state fMRI with AM-FM func-
tions as follows: First, we compute the AM-FM decomposition for each subject
4-dimensional data, from which we extract the instantaneous frequency over time
(IFt) and instantaneous frequency over space (IFs =√IF 2
x + IF 2y + IF 2
z ); then, I
create a bivariate histogram of IFt and IFs) with 20 bins for each feature and for
each patient.
Each histogram summarizes the behavior of a subject in resting state. Thus, we
would expect to find evidence of differences between patients and healthy controls
by a comparison of these bivariate histograms. We proceed to compare each bin
across groups to establish differences between patients and controls with a logistic
regression. The regression model is built using a case or control indicator function
as response and the bin proportion, age and gender as explanatory variables. Fig.
3.4 shows an example bivariate histogram for one healthy subject.
3.2.3 Instantaneous frequency analysis of sMRI
Structural MRI reveals brain anatomy (gray matter concentration) in higher spatial
resolution than functional MRI. A natural approach is to extract texture information
from gray matter concentration, which provides a description of the variation in
intensity, including spatial patterns imperceptible to the naked eye. Texture analysis
on MRI has proven successful for abnormality detection [56, 57]. Moreover, AM-FM
decomposition have also proven useful for the study of lesions in multiple sclerosis
patients [58].
29
Chapter 3. Preliminary results from AM-FM analysis of functional and structural MRI
The gray matter concentration value is subject to scanner variability either in
multisite studies and even in the same scanner over time [59, 60]. On the other hand,
the patterns of intensity variation are more likely to be consistent across different
scanners with a consistent scanning protocol [61].
The strategy to compare between patients and healthy controls in the sMRI
setting will be to obtain the IFS for each voxel and use this information to test
between groups using voxel-wise two sample t-tests with family-wise error correction.
Notice the for sMRI IFt is not defined since no time information is available in
this case.
30
Chapter 3. Preliminary results from AM-FM analysis of functional and structural MRI
Figure 3.2: Matrix plot of AM-FM features and BOLD fMRI for raw data (upperplot) and after edge and eye blinking artifacts have been removed (lower plot).
31
Chapter 3. Preliminary results from AM-FM analysis of functional and structural MRI
(a) (b)
Figure 3.3: (a) Manually selected Principal component showing coherence with brainfunction using the feature space containing AM-FM functions and (b) coefficients ofthe component corresponding to IP and expected response (HRF) of the BOLDsignal for the task exiciing the right visual and motor cortices.
(a) (b)
Figure 3.4: (a) Scatter plot of IFS and IFt and (b) bivariate histogram for one healthysubject. A logarithm transformation was applied for the probability frequencies ineach bin to enhance visualization.
32
Chapter 4
Results
In the previous chapter, we suggested methods to analyze of AM-FM features of
applied task-related fMRI, resting-state fMRI and structural MRI (sMRI). The goals
were as follows:
Task fMRI: Identify brain regions that are biologically meaningful and whose tem-
poral activity is in response to task stimulus.
Resting state fMRI: Use AM-FM features to detect abnormal activity in patients,
as compared to healthy controls.
Structural MRI: Capture structural changes in gray-matter between patients and
healthy controls.
In the present chapter, we show results of each method and discuss the advantages
of applying this new form of data representation. First, we present the results of ana-
lyzing IP features from task-related fMRI. Then, we compare the bivariate histogram
representations from resting-state fMRI of schizophrenia patients and healthy con-
trols. Finally, we perform a mass univariate two sample test on the voxel-wise mod-
33
Chapter 4. Results
(a) (b) (c)
Figure 4.1: Instantaneous phase of axial slice 11 (left) and slice 24 (right) to showactivation at visual and motor areas in the brain for a subject (a) in rest, (b) excitingleft side and (c) right side of the brain.
ulo of the instantaneous frequency between patients with schizophrenia and healthy
controls.
4.1 Task-related fMRI analysis
The patient performed a task that simultaneously excited visual and motor cortices
in the brain. The scanning protocol of this subject is described in section 3.2. The
data consist of a four dimensional hypercube of dimensions 53× 63× 46× 117.
In section 3.2, we concluded that the IP is the most informative AM-FM feature.
In the following, we further explore the properties of the instantaneous phase.
The IP responds to brain regions that are active, with a pattern in time that is
coherent with the theoretical HRF-convolved stimulus. A visual inspection of the
axial slices on the instantaneous phase in Fig. 4.1 shows it is sufficient to detect
activation of visual and motor areas in the brain. Here, we note that this activation
is not noticeable to the naked eye in the raw data.
The IP also reveals contrast in areas of the brain where we expect to show activity.
The response in these areas is captured in form of de-phasing from the rest of the
34
Chapter 4. Results
(a) (b)
Figure 4.2: (a) Histogram of the instantaneous phase and (b) Histogram of the rawfMRI data.
brain. Therefore, we propose the use of the IP for detecting task-activated brain
regions.
The distribution of the IP is highly kurtodic and has zero mean as shown in
Fig. 4.2(a). This makes the IP adequate for a spatial ICA decomposition using the
Infomax algorithm with sigmoid non-linearity and, thus, we exlpore independent IP
components. In contrast, the distribution of the original data presents more Gaussian
and sub-Gaussian features, like strong skewness and low kurtosis.
Next, we present the results of applying ICA to the IP feature (Fig. 4.3). Com-
pared to the activation maps with better clustering obtained from the use of ICA on
the original fMRI signal, ICA on IP provides maps of the visual and motor cortices,
and, particularly, less granularity.
35
Chapter 4. Results
(a) (b)
Figure 4.3: Biologically relevant ICA on instantaneous phase components for (a) lefttask and (b) right task.
4.2 Resting-state fMRI analysis
In this case, the data set consists of 369 subjects scanned while at rest. Among
them, 180 have been diagnosed with schizophrenia while 189 were healthy controls.
We compute the bivariate histogram of IFs vs IFt for each individual as previously
described in sec. 3.2.2. This provides a unique representation for each subject.
Then, we fit a logistic regression model for each bin of the bivariate histogram us-
ing the following explanatory variables: the proportion value of the bin, age, gender
and diagnosis as response variable. We also performed a backwards model selection
procedure with the interactions. Then, we evaluated each model and dropped in-
teraction terms that were not significant, i.e., with the significance level above 0.1.
In Fig. 4.4, we show the bins for which coefficients were significant at the p = 0.01
level after Bonferroni correction (0.01/400). The figure shows evidence of differences
36
Chapter 4. Results
(a) (b)
Figure 4.4: (a) P-value map for each bin tested and (b) bins that passed Bonferronicorrection.
in certain, frequency bins, which we recommend as good features for automated
diagnosis (classification) in the future.
4.3 Structural MRI analysis
Here, the data set consists on sMRI from 353 subjects. Among them, 165 are diag-
nosed with schizophrenia and 188 are healthy controls. In order to identify regions
that are different between schizophrenic patients and healthy controls, we compare
the IF magnitude at each voxel using a two sample t-test (significance level of 0.01
after Bonferroni correction).
From the results in Fig. 4.5, we can observe which voxels have higher mean IF
magnitude in the patients, as compared to the healthy controls. This is evidence
that the IF magnitude over a small brain region can be used to differentiate cases of
schizophrenia from healthy controls (see Fig. 4.5).
37
Chapter 4. Results
Figure 4.5: log10(p) thresholded at 5 with a T1 image as template.
38
Chapter 5
Conclusion and future work
In this chapter, we provide final conclusions based on the results from the previous
chapter and suggest future work.
5.1 Conclusion
In the thesis, the focus was to introduce the use of AM-FM analysis methods to MRI
images. For analyzing MRI images, we first proposed a comparison framework for
selecting the best combination of an AM-FM method and filterbank configuration. In
terms of the mean square error of the IF magnitude estimation from a synthetic FM
image. We found that the combination of the multi-dimensional quasi-local method
with an equiripple filter-bank gave most accurate and robust to noise estimation.
The application of the quasi-local method with an equiripple filterbank gave the
following results:
Structural MRI: The IF magnitude associated with gray matter concentration
showed that select, local regions of the brain exhibited significant difference
39
Chapter 5. Conclusion and future work
between patients diagnosed with schizophrenia and healthy controls.
Resting fMRI: The joint-histogram of the IF spatial frequency magnitude (||(φx, φy)||)
and the IF-time frequency component (|φt|) provided for an effective charac-
terization of brain activity. We fit logistic regression for each bin proportion
value as a regressor along with age and gender showing evidence of abnormal
proportions for one bin in schizophrenia patients.
Task fMRI: The instantaneous phase (IP) showed response to brain activity by
identifying the regions that get out of phase compared to the rest of the brain.
AM-FM models has been applied in sMRI images for analysis of brain lesions
in [58], although, to the best of our knowledge, this thesis is the first approach of
AM-FM models to fMRI.
5.2 Future work
This section provides ideas for further research in this topic. Future work can address
the following issues:
• Implementation of reconstruction from IP and IA on fMRI after ignoring IP
values for low IA.
• Adaptive multi-scale AM-FM demodulation.
• Combination of ICA and IP can be studied as separation of multiplicative
sources.
• Filterbank design for multicomponent analysis.
• Use of ICA for separating multiple components from a single channel.
40
Appendix A
Proof that ω(x) ∈ [0, π3 ] when
R ∈ [−∞,−1]
Recall from sec. 1.2.2 that ω(x) = cos−1 θ(R) and θ(R) =
[R +√R2 + 8
4
]
First, notice that θ(R) is a monotonically increasing function.
Let x1 > x2, ∀x1, x2 ∈ R
x1 − x2 > 0 (A.1)√x21 + 8−
√x22 + 8 > 0 (A.2)
Now, adding eq. (A.1) and eq. (A.2)
x1 +√x21 + 8− x2 −
√x22 + 8 > 0
θ(x1)− θ(x2) > 0
41
Appendix A. Proof that ω(x) ∈ [0, π3] when R ∈ [−∞,−1]
Then:
−∞ < R < −1
θ(−∞) < θ(R) < θ(−1)
where
θ(−∞) = limR→−∞
[R +√R2 + 8
4
]
= limR→−∞
[R +√R2 + 8
4
]× (R−
√R2 + 8)
(R−√R2 + 8)
= limR→−∞
R2 − (R2 + 8)
4(R−√R2 + 8)
= limR→−∞
2√R2 + 8−R
=2√
(−∞)2 + 8− (−∞)
=2
∞= 0
θ(−1) =−1 +
√(−1)2 + 8
4
= 0.5
Therefore: 0 < θ(R) < 0.5 and according to eq. (1.3) 0 < ω(x) < π3.
42
Appendix B
Proof that ω(x) ∈ [2π3 , π] when
R ∈ [1,∞]
Recall from sec. 1.2.2, that ω(x) = π − cos−1 θ(R) and θ(R) =
[−R +
√R2 + 8
4
]
First, notice that θ(R) is a monotonically decreasing function.
We have:
dθ(R)
dR< 0
−1
4+
R
4√R2 + 8
< 0
R√R2 + 8
< 1
R <√R2 + 8
R2 < R2 + 8
0 < 8
43
Appendix B. Proof that ω(x) ∈ [2π3, π] when R ∈ [1,∞]
Then:
1 < R <∞
and
θ(1) > θ(R) > θ(∞)
where
θ(∞) = limR→−∞
[−R +
√R2 + 8
4
]
= limR→−∞
[−R +
√R2 + 8
4
]× (−R−
√R2 + 8)
(−R−√R2 + 8)
= limR→−∞
R2 − (R2 + 8)
4(−R−√R2 + 8)
= limR→−∞
2√R2 + 8 +R
=2√
(∞)2 + 8 + (∞)
=2
∞= 0
θ(1) =−1 +
√(1)2 + 8
4
= 0.5
Therefore, 0 < θ(R) < 0.5 and according to eq. (1.5) 2π3< ω(x) < π.
44
References
[1] A. L. da Cunha, J. Zhou, and M. N. Do, “The nonsubsampled contourlet trans-form: theory, design, and applications.,” IEEE transactions on image process-ing: a publication of the IEEE Signal Processing Society, vol. 15, pp. 3089–101,octubre 2006.
[2] M. N. Do, Directional multiresolution image representations. PhD thesis, poly-technique Federale De Lausanne, E., 2001.
[3] V. Calhoun, T. Adali, G. Pearlson, and J. Pekar, “Group ica of functional mridata: separability, stationarity, and inference,” in Proc. Int. Conf. on ICA andBSS San Diego, CA. p, vol. 155, 2001.
[4] S. Kisner and T. Talavage, “Testing the distribution of nonstationary mri data,”in Engineering in Medicine and Biology Society, 2004. IEMBS ’04. 26th AnnualInternational Conference of the IEEE, vol. 1, pp. 1888 –1891, sept. 2004.
[5] C. Long, E. Brown, C. Triantafyllou, I. Aharon, L. Wald, V. Solo, et al., “Non-stationary noise estimation in functional mri.,” NeuroImage, vol. 28, no. 4,p. 890, 2005.
[6] L. Cohen, Time-frequency analysis, vol. 778. Prentice Hall PTR EnglewoodCliffs, New Jersey, 1995.
[7] I. Daubechies, “The wavelet transform, time-frequency localization and signalanalysis,” Information Theory, IEEE Transactions on, vol. 36, pp. 961 –1005,sep 1990.
[8] J. F. Kaiser, “On a simple algorithm to calculate the ‘energy’ of a signal,” inAcoustics, Speech, and Signal Processing, 1990. ICASSP-90., 1990 InternationalConference on, pp. 381 –384 vol.1, apr 1990.
45
References
[9] J. F. Kaiser, “On Teager’s energy algorithm and its generalization to continuoussignals,” in Proceedings of IEEE DSP Workshop, (New Paltz, NY), Sept. 1990.
[10] J. Havlicek, AM-FM image models. PhD thesis, The University of Texas Austin,1996.
[11] M. Pattichis et al., AM-FM transforms with applications. PhD thesis, TheUniversity of Texas at Austin, 1998.
[12] M. Pattichis and A. Bovik, “Analyzing image structure by multidimensional fre-quency modulation,” Pattern Analysis and Machine Intelligence, IEEE Trans-actions on, vol. 29, pp. 753 –766, may 2007.
[13] V. Murray, AM-FM methods for image and video processing. PhD thesis, Uni-versity of New Mexico, 2008.
[14] K. Kwong, J. Belliveau, D. Chesler, I. Goldberg, R. Weisskoff, B. Poncelet,D. Kennedy, B. Hoppel, M. Cohen, and R. Turner, “Dynamic magnetic res-onance imaging of human brain activity during primary sensory stimulation,”Proc. Natl. Acad. Sci., vol. 89, pp. 5675–5679, 1992.
[15] D. Noll, “A primer on mri and functional mri,” tech. rep., University of Michi-gan, Ann Arbor, USA, 2001.
[16] D. I. Hoult, C.-N. Chen, and V. J. Sank, “Quadrature detection in the laboratoryframe,” Magnetic Resonance in Medicine, vol. 1, no. 3, pp. 339–353, 1984.
[17] Institute of Neurology, UCL, 12 Queen Square, London WC1N 3BG, UK, SPM8Manual, February 2012.
[18] M. S. Pattichis, G. Panayi, A. Bovik, and and H. Shun-Pin, “Fingerprint clas-sification using and am-fm model,” IEEE Transactions on Image Processing,vol. 10, pp. 951–954, 2001.
[19] N. Ray, J. Havlicek, S. T. Acton, and M. S. Pattichis, “Active contour seg-mentation guided by am-fm dominant component analysis,” in in Proc. IEEEInternational Conference on Image Processing, pp. 78–81, 2001.
[20] T. B. Yap, J. P. Havlicek, and V. DeBrunner, “Bayesian segmentation of am-fmtexture images,” in Proc. 35th Asilomar Conf. Signals, Syst., Comput., vol. 2,pp. 1156–1160, 2001.
[21] M. S. Pattichis, C. S. Pattichis, M. Avraam, A. Bovik, and K. Kyriakou, “Am-fm texture segmentation in electron microscopic muscle imaging,” Acoustics,Speech, and Signal Processing, vol. 4, pp. 2331–2334, 1999.
46
References
[22] P. Rodriguez, M. S. Pattichis, and M. B. Goens, “M-mode echocardiographyimage and video segmentation based on am-fm demodulation techniques,” inin International Conference of the IEEE Engineering in Medicine and BiologySociety, 2003.
[23] C. Christodoulou, C. Pattichis, V. Murray, M. Pattichis, and A. Nicolaides,“AM-FM Representations for the Characterization of Carotid Plaque Ultra-sound Images,” in 4th European Conference of the International Federation forMedical and Biological Engineering, pp. 459–546, Springer Berlin Heidelberg,2008.
[24] P. Rodriguez, Fast and Accurate AM-FM demodulation of Digital Images WithApplications. PhD thesis, The University of New Mexico, 2005.
[25] S. Murillo, M. Pattichis, C. Loizou, C. Pattichis, E. Kyriacou, A. Constantinides,and A. Nicolaides, “Atherosclerotic Plaque Motion Analysis from UltrasoundVideos,” in Fortieth Asilomar Conference on Signals, Systems and Computers,2006. ACSSC’06., pp. 836–840, 2006.
[26] J. Havlicek and A. Bovik, Image modulation models, ch. 4, pp. 313–324. Burling-ton: Elsevier Academic Press, 2000.
[27] V. Murray, P. Rodriguez, and M. Pattichis, “Multi-scale AM-FM Demodulationand Image Reconstruction Methods with Improved Accuracy,” IEEE transac-tions on image processing, 2010.
[28] M. Martin, “Statistical analysis of fmri time-series: A critical review of the glmapproach,” Frontiers in Human Neuroscience, vol. 5, no. 00028, 2011.
[29] V. D. Calhoun, T. Adali, L. K. Hansen, J. Larsen, and J. J. Pekar, “Ica of func-tional mri data: An overview,” in in Proceedings of the International Workshopon Independent Component Analysis and Blind Signal Separation, pp. 281–288,2003.
[30] V. D. Calhoun, T. Adali, and J. J. Pekar, “A method for comparing group fmridata using independent component analysis: application to visual, motor andvisuomotor tasks,” Magnetic Resonance Imaging, vol. 22, pp. 1181–1191, 2004.
[31] A. Hyvarinen and E. Oja, “Independent component analysis: algorithms andapplications,” Neural Networks, vol. 13, pp. 411 – 430, 2000.
[32] T. Lee, Independent Component Analysis: Theory and Applications. Springer,1998.
47
References
[33] K. V. Mardia, “Applications of some measures of multivariate skewness andkurtosis in testing normality and robustness studies,” Sankhya: The IndianJournal of Statistics, Series B (1960-2002), vol. 36, no. 2, pp. pp. 115–128,1974.
[34] L. Brillouin, “The negentropy principle of information,” Journal of AppliedPhysics, vol. 24, pp. 1152 –1163, sep 1953.
[35] A. Hyvarinen, “Independent component analysis by minimization of mutualinformation,” 1997.
[36] A. Bell and T. Sejnowski, “An information-maximization approach to blindseparation and blind deconvolution,” Neural computation, vol. 7, no. 6, pp. 1129–1159, 1995.
[37] D. Pham and P. Garat, “Blind separation of mixture of independent sourcesthrough a maximum likelihood approach,” in In Proc. EUSIPCO, Citeseer, 1997.
[38] V. Calhoun, T. Adali, G. Pearlson, and J. Pekar, “Spatial and temporal indepen-dent component analysis of functional mri data containing a pair of task-relatedwaveforms,” Human brain mapping, vol. 13, no. 1, pp. 43–53, 2001.
[39] N. C. Srinivas Rachakonda, Eric Egolf and V. Calhoun, Group ICA Of fMRIToolbox Manual. The Mind Research Network, August 2004.
[40] A. Bovik, “Variational pattern analysis using gabor wavelets,” in ICASSP-92:1992 IEEE International Conference on Acoustics, Speech, and Signal Process-ing, 1992.
[41] M. C. A. C. Bovik and W. S. Geisler, “Computational texture analysis usinglocalized spatial filtering,” in IEEE Computer Society Workshop Computer Vi-sion, 1987.
[42] A. Bovik and M. Clark, “Multichannel texture analysis using localized spatialfilters,” Transactions on Pattern Analysis, 1990.
[43] M. Clark, “Texture discrimination using a model of visual cortex,” in Proc.IEEE Conf. on Systems, Man and Cybernetics, 1986.
[44] J.-H. Chang, K.-C. Fan, and Y.-L. Chang, “Multi-modal gray-level histogrammodeling and decomposition,” Image and Vision Computing, vol. 20, no. 3,pp. 203 – 216, 2002.
48
References
[45] B. Biswal, F. Zerrin Yetkin, V. Haughton, and J. Hyde, “Functional connectivityin the motor cortex of resting human brain using echo-planar mri,” Magneticresonance in medicine, vol. 34, no. 4, pp. 537–541, 1995.
[46] M. Greicius, B. Flores, V. Menon, G. Glover, H. Solvason, H. Kenna, A. Reiss,and A. Schatzberg, “Resting-state functional connectivity in major depression:abnormally increased contributions from subgenual cingulate cortex and thala-mus,” Biological psychiatry, vol. 62, no. 5, pp. 429–437, 2007.
[47] D. Kennedy, E. Redcay, and E. Courchesne, “Failing to deactivate: restingfunctional abnormalities in autism,” Proceedings of the National Academy ofSciences, vol. 103, no. 21, pp. 8275–8280, 2006.
[48] L. Tian, T. Jiang, Y. Wang, Y. Zang, Y. He, M. Liang, M. Sui, Q. Cao, S. Hu,M. Peng, et al., “Altered resting-state functional connectivity patterns of an-terior cingulate cortex in adolescents with attention deficit hyperactivity disor-der,” Neuroscience letters, vol. 400, no. 1-2, pp. 39–43, 2006.
[49] M. Liang, Y. Zhou, T. Jiang, Z. Liu, L. Tian, H. Liu, and Y. Hao, “Widespreadfunctional disconnectivity in schizophrenia with resting-state functional mag-netic resonance imaging,” Neuroreport, vol. 17, no. 2, pp. 209–213, 2006.
[50] C. Beckmann, M. DeLuca, J. Devlin, and S. Smith, “Investigations into resting-state connectivity using independent component analysis,” Philosophical Trans-actions of the Royal Society B: Biological Sciences, vol. 360, no. 1457, pp. 1001–1013, 2005.
[51] M. Greicius, B. Krasnow, A. Reiss, and V. Menon, “Functional connectivity inthe resting brain: a network analysis of the default mode hypothesis,” Proceed-ings of the National Academy of Sciences, vol. 100, no. 1, pp. 253–258, 2003.
[52] A. Di Martino, A. Scheres, D. Margulies, A. Kelly, L. Uddin, Z. Shehzad,B. Biswal, J. Walters, F. Castellanos, and M. Milham, “Functional connec-tivity of human striatum: a resting state fmri study,” Cerebral cortex, vol. 18,no. 12, pp. 2735–2747, 2008.
[53] E. Allen, E. Erhardt, E. Damaraju, W. Gruner, J. Segall, R. Silva, M. Havlicek,S. Rachakonda, J. Fries, R. Kalyanam, et al., “A baseline for the multivariatecomparison of resting-state networks,” Frontiers in systems neuroscience, vol. 5,2011.
[54] D. Cordes, V. Haughton, K. Arfanakis, J. Carew, P. Turski, C. Moritz,M. Quigley, and M. Meyerand, “Frequencies contributing to functional connec-tivity in the cerebral cortex in resting-state data,” American Journal of Neuro-radiology, vol. 22, no. 7, pp. 1326–1333, 2001.
49
References
[55] P. Fransson, “Spontaneous low-frequency bold signal fluctuations: An fmri in-vestigation of the resting-state default mode of brain function hypothesis,” Hu-man brain mapping, vol. 26, no. 1, pp. 15–29, 2005.
[56] J. Zhang, L. Tong, L. Wang, and N. Li, “Texture analysis of multiple sclerosis: acomparative study,” Magnetic resonance imaging, vol. 26, no. 8, pp. 1160–1166,2008.
[57] D. Mahmoud-Ghoneim, G. Toussaint, J. Constans, and J. de Certaines, “Threedimensional texture analysis in mri: a preliminary evaluation in gliomas,” Mag-netic resonance imaging, vol. 21, no. 9, pp. 983–987, 2003.
[58] C. Loizou, V. Murray, M. Pattichis, I. Seimenis, M. Pantziaris, and C. Pat-tichis, “Multiscale amplitude-modulation frequency-modulation (am–fm) tex-ture analysis of multiple sclerosis in brain mri images,” Information Technologyin Biomedicine, IEEE Transactions on, vol. 15, no. 1, pp. 119–129, 2011.
[59] V. Gradin, V.-E. Gountouna, G. Waiter, T. S. Ahearn, D. Brennan, B. Condon,I. Marshall, D. J. McGonigle, A. D. Murray, H. Whalley, J. Cavanagh, D. Hadley,K. Lymer, A. McIntosh, T. W. Moorhead, D. Job, J. Wardlaw, S. M. Lawrie, andJ. D. Steele, “Between- and within-scanner variability in the calibrain study n-back cognitive task,” Psychiatry Research: Neuroimaging, vol. 184, no. 2, pp. 86– 95, 2010.
[60] V.-E. Gountouna, D. E. Job, A. M. McIntosh, T. W. J. Moorhead, G. K. L. Ly-mer, H. C. Whalley, J. Hall, G. D. Waiter, D. Brennan, D. J. McGonigle, T. S.Ahearn, J. Cavanagh, B. Condon, D. M. Hadley, I. Marshall, A. D. Murray,J. D. Steele, J. M. Wardlaw, and S. M. Lawrie, “Functional magnetic resonanceimaging (fmri) reproducibility and variance components across visits and scan-ning sites with a finger tapping task,” NeuroImage, vol. 49, no. 1, pp. 552 – 560,2010.
[61] G. Collewet, M. Strzelecki, and F. Mariette, “Influence of mri acquisition pro-tocols and image intensity normalization methods on texture classification,”Magnetic Resonance Imaging, vol. 22, no. 1, pp. 81 – 91, 2004.
50